X2Edit: Revisiting Arbitrary-Instruction Image Editing through Self-Constructed Data and Task-Aware Representation Learning
Jian Ma¹, Xujie Zhu², Zihao Pan², Qirong Peng¹, Xu Guo³, Chen Chen¹, Haonan Lu¹

¹OPPO AI Center, ²Sun Yat-sen University, ³Tsinghua University

X2Edit image generation results

News

• 2025/09/16: We release a dataset built using Qwen-Image and Qwen-Image-Edit. This sub-dataset specifically focuses on subject-driven generation with facial consistency—a key requirement for tasks requiring stable subject identity across generated content. Asian-portrait and NonAsian-portrait
• 2025/08/25: Support Qwen-Image for training and inference. Checkpoint

X2Edit image generation results with Qwen-Image

Environment

Prepare the environment, install the required libraries:

$ cd X2Edit
$ conda create --name X2Edit python==3.11
$ conda activate X2Edit
$ pip install -r requirements.txt

Clone LaMa to data_pipeline and rename it to lama. Clone SAM and GroundingDINO to SAM, and then rename them to segment_anything and GroundingDINO

Data Construction

(./assets/dataset_detail.jpg)

X2Edit provides executable scripts for each data construction workflow shown in the figure. We organize the dataset using the WebDataset format. Please replace the dataset in the scripts. The following Qwen model can be selected from Qwen2.5-VL-72B, Qwen3-8B, and Qwen2.5-VL-7B. In addition, we also use aesthetic scoring models for screening, please donwload SigLIP and aesthetic-predictor-v2-5, and then change the path in siglip_v2_5.py.

• Subject Addition & Deletion → use expert_subject_deletion.py and expert_subject_deletion_filter.py: The former script is used to construct deletion-type data, while the latter uses the fine-tuned Qwen2.5-VL-7B to further screen the constructed deletion-type data. Before executing, download RAM, GroundingDINO, SAM, Randeng-Deltalm, InfoXLM, RMBG and LaMa.
• Normal Editing Tasks → use step1x_data.py: Please download the checkpoint Step1X-Edit. The language model we use is Qwen2.5-VL-72B.
• Subject-Driven Generation → use kontext_subject_data.py: Please download the checkpoints FLUX.1-Kontext, DINOv2, CLIP, OPUS-MT-zh-en, shuttle-3-diffusion. The language model we use is Qwen3-8B.
• Style Transfer → use kontext_style_transfer.py: Please download the checkpoints FLUX.1-Kontext, DINOv2, CLIP, OPUS-MT-zh-en, shuttle-3-diffusion. The language model we use is Qwen3-8B.
• Style Change → use expert_style_change.py: Please download the checkpoints FLUX.1-dev, OmniConsistency. We use Qwen2.5-VL-7B to score.
• Text Change → use expert_text_change_ch.py for Chinese and use expert_text_change_en.py for English: Please download the checkpoint textflux. We use Qwen2.5-VL-7B to score.
• Complex Editing Tasks → use bagel_data.py: Please download the checkpoint Bagel. We use Qwen2.5-VL-7B to score.
• High Fidelity Editing Tasks → use gpt4o_data.py: Please download the checkpoint OPUS-MT-zh-en and use your own GPT-4o API. We use Qwen2.5-VL-7B to score.
• High Resoluton Data Construction → use kontext_data.py: Please download the checkpoint FLUX.1-dev and OPUS-MT-zh-en. We use Qwen2.5-VL-7B to score.

Inference

We provides inference scripts for editing images with resolutions of 1024 and 512. In addition, we can choose the base model of X2Edit, including FLUX.1-Krea, FLUX.1-dev, FLUX.1-schnell, PixelWave, shuttle-3-diffusion, and choose the LoRA for integration with MoE-LoRA including Turbo-Alpha, AntiBlur, Midjourney-Mix2, Super-Realism, Chatgpt-Ghibli. Choose the model you like and download it. For the MoE-LoRA, we will open source a unified checkpoint that can be used for both 512 and 1024 resolutions.

Before executing the script, download Qwen3-8B to select the task type for the input instruction, base model(FLUX.1-Krea, FLUX.1-dev, FLUX.1-schnell, shuttle-3-diffusion), MLLM and Alignet. All scripts follow analogous command patterns. Simply replace the script filename while maintaining consistent parameter configurations.

$ python infer.py --device cuda --pixel 1024 --num_experts 12 --base_path BASE_PATH --qwen_path QWEN_PATH --lora_path LORA_PATH --extra_lora_path EXTRA_LORA_PATH
$ python infer_qwen.py --device cuda --pixel 1024 --num_experts 12 --base_path BASE_PATH --qwen_path QWEN_PATH --lora_path LORA_PATH --extra_lora_path EXTRA_LORA_PATH  ## for Qwen-Image backbone

device: The device used for inference. default: cuda
pixel: The resolution of the input image, , you can choose from [512, 1024]. default: 1024
num_experts: The number of expert in MoE. default: 12
base_path: The path of base model.
qwen_path: The path of model used to select the task type for the input instruction. We use Qwen3-8B here.
lora_path: The path of MoE-LoRA in X2Edit.
extra_lora_path: The path of extra LoRA for plug-and-play. default: None.

Train

We organize the dataset using the WebDataset format. Please replace the dataset in the training script. Before executing, download FLUX.1-dev, MLLM and Alignet, replace the paths in train_1024.sh and train_512.sh. Then you can run:

• For 1024 resolution

  bash training_script/train_1024.sh
  bash training_script/train_qwen_x2edit_1024.sh  ## for Qwen-Image backbone
  

• For 512 resolution

  bash training_script/train_512.sh
  bash training_script/train_qwen_x2edit_512.sh  ## for Qwen-Image backbone
  

Important parameters in script
rank: The rank in LoRA. default: 64
experts: The number of expert in MoE. default: 12
task: The number of editing task. default: 16
X2I_MLLM: The checkpoint path of MLLM.
X2I_alignet: The checkpoint path of Alignet.

Evaluation

We provides evaluation scripts for the calculation of VIEScore and ImgEdit-Judge-Score. Ensure that the edited images and the original images are placed in the same folder and named using the format "instruction + suffix". For example:

original image: Change the background to the sea._in.jpg
edited image: Change the background to the sea._out.jpg

in this case, instruction is "Change the background to the sea", suffix of original image is "_in.jpg", suffix of edited image is "_out.jpg".

• VIEScore(for gpt4o or qwen25vl):

$ cd evaluation
$ python score_i2i_new.py --backnone gpt4o --in_suffix _in.jpg --out_suffix _out.jpg --ref_dir YOUR_DATA_PATH --output_txt_dir ./score

• ImgEdit-Judge-Score:

$ cd evaluation
$ python score_i2i_new.py --backnone ImgEditJudge --in_suffix _in.jpg --out_suffix _out.jpg --ref_dir YOUR_DATA_PATH --output_txt_dir ./score

backnone: Model for evaluation, you can choose from ['gpt4o', 'qwen25vl', 'ImgEditJudge']. For qwen25vl, download Qwen2.5-VL-72B and change the path in __init__.py. For ImgEditJudge, download the fine-tuned Qwen2.5-VL-7B.
in_suffix: The suffix of the original image file name.
out_suffix: The suffix of the edited image file name.
ref_dir: The folder of edited images and original images.
output_txt_dir: The folder for file used to record scores.

• Calculate Average Score:

$ cd evaluation
$ python calculate_score_en_ch.py

Acknowledgements

This code is built on the code from the diffusers and EasyControl.

Citation

🌟 If you find our work helpful, please consider citing our paper and leaving valuable stars

@misc{ma2025x2editrevisitingarbitraryinstructionimage,
      title={X2Edit: Revisiting Arbitrary-Instruction Image Editing through Self-Constructed Data and Task-Aware Representation Learning}, 
      author={Jian Ma and Xujie Zhu and Zihao Pan and Qirong Peng and Xu Guo and Chen Chen and Haonan Lu},
      year={2025},
      eprint={2508.07607},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.07607}, 
}